Fast CUDA SGEMM from Scratch

Step-by-step optimization of matrix multiplication, implemented in CUDA. For an explanation of each kernel, see siboehm.com/CUDA-MMM. This repo is inspired by wangzyon/NVIDIA_SGEMM_PRACTICE.

Overview

Running the kernels on a NVIDIA A100 (Ampere):

GFLOPs at matrix size 4092x4092:

Kernel	GFLOPs/s	Performance relative to cuBLAS
1: Naive	292	1.7%
2: GMEM Coalescing	3115.7	17.8%
3: SMEM Caching	5448.6	31.1%
4: 1D Warptiling	10345.5	59.0%
5: 2D Warptiling	14126.6	80.6%
8: Avoid Bank Conflicts (Offset)	15056.9	85.9%
7: Avoid Bank Conflicts (Linearize)	15157.5	86.5%
6: Vectorized Mem Access	15334.9	87.5%
9: Autotuning	15664.8	89.4%
0: cuBLAS	17521.2	100.0%

Setup

1. Install dependencies: CUDA toolkit, Python (+ Seaborn), CMake, Ninja. See environment.yml.
2. Configure NVCC compilation parameters. Look up your GPUs compute capability here. Then configure the CMakeLists.txt and change:

   set(CUDA_COMPUTE_CAPABILITY 80)

3. Build: make
4. Run one of the kernels: DEVICE=<device_id> ./sgemm <kernel number>
5. Profiling via NVIDIA Nsight Compute (ncu): make profile KERNEL=<kernel number>